Column

Agent Engineering

Five essays from context foundations to production systems

5 essays · 114 min total

This column traces one complete path: from principles, to practice, to reflection.

The first essay lays the foundation — context is not prompt, and why context engineering is becoming the new bedrock of AI applications. Every architectural decision that follows builds on this layer.

The second opens the panorama — where the 98.4% of agent engineering actually lives. The model is just 1.6%; the rest is harness.

The third and fourth are two sharply contrasting case studies: open-lovable tames bare APIs without any agent framework, while Relay is a production job-hunting agent system designed from scratch. One subtracts; the other adds.

The final essay steps outside engineering — when agents gain memory and continuity, where does the “self” begin? From Locke’s personal identity to OpenClaw, philosophy and engineering look each other in the eye.

Reading 1 → 5 in order works best, but every essay stands on its own.

Contents

  1. Dissecting open-lovable: An App Generator That Tames the Raw API Without an Agent Framework

    A full dissection of firecrawl/open-lovable (27k★, paste a URL and get a working React app in seconds), from product to code. Its most interesting trait isn't that it generates code — it's that it uses no agent framework, no Claude Agent SDK, no native tool-calling. Instead it hand-rolls an entire harness on top of the raw LLM API: a text DSL protocol, streaming regex parsing, truncation detection and recovery, manual context orchestration, plus a swappable cloud sandbox layer (E2B / Vercel Sandbox). This is a case study in taming the raw API.

    · 26 min

  2. Building a Production-Grade AI Agent System from Scratch: A Full Architecture Breakdown of Relay

    Using the Relay open-source job-search Agent project as a case study, this article fully breaks down every key design decision in a production-grade multi-agent system: why split a single Agent into 5, how to implement HITL checkpoints with LangGraph, how a three-tier LLM router precisely tracks costs, how a fabrication guard validates at runtime, and how a hybrid backend (Hono/Bun + FastAPI/Python) decouples cleanly. Whether you are building your first Agent PoC or pushing toward production, there are design patterns here you can take away.

    · 20 min

  3. Context Is Not Prompt: Why Context Engineering Is Becoming AI's New Foundation

    Why context engineering supersedes prompt engineering — a systematic look at context assembly, retrieval, compression, and eviction patterns, drawing from Anthropic, Karpathy, LangChain, and Manus.

    · 16 min

  4. The Agent Engineering Map: Where Does That 98.4% of the Work Actually Live?

    A panoramic map that treats Agent Engineering as a discipline. Starting from the widely cited claim that only 1.6% of Claude Code is AI decision logic while 98.4% is infrastructure, it walks the eight pillars one by one — orchestration, context, memory, tools, reliability, evaluation, cost, governance — explaining the gap each fills, its minimal implementation, and its failure boundary. It fuses 2025 to 2026 frontline engineering from Anthropic, OpenAI, Cognition, Manus, and Temporal, and lands on one line: the model is bought, the harness is built, and your entire engineering leverage lives in that 98.4%.

    · 30 min

  5. Agent Identity: From Locke to OpenClaw

    AI agent amnesia isn't a functional defect—it's a fundamental gap in the trust account. Starting from Locke's 1689 theory of identity, this article dissects the complete engineering stack for agent identity continuity in 2026: file-as-identity (SOUL.md paradigm), Harness as environmental condition, four-layer memory architecture and Gene Capsule protocol, self-positioning in multi-agent topology, and evaluation as the ultimate identity verification challenge. For practitioners building or designing AI agent systems, and researchers deeply thinking about the boundaries of AI autonomy.

    · 22 min